Google Fortifies Chrome Agentic AI Against Indirect Prompt Injection Attacks

On 8 December 2025, SecurityWeek reported that Google is strengthening Chrome’s agentic browsing protections to counter indirect prompt injection attacks that could lead to data leaks or unwanted actions. According to Google, a new AI model built with Gemini, called the User Alignment Critic, is isolated from untrusted content to vet the agent’s actions and veto any misaligned ones, while only metadata about the proposed action is visible to it.

Chrome’s Site Isolation and same-origin policy protections are being expanded with Agent Origin Sets, designed to prevent a compromised agent from acting arbitrarily on unrelated origins. A gating function separates relevant origins into read-only and read-write categories, limiting the model’s exposure to cross-site data and checking each new origin for relevancy before navigation.

The system also logs work transparently, and requires user confirmation before sensitive actions such as visiting certain sites, signing in, making purchases, or sending messages, with a prompt-injection classifier running in parallel to stop actions tied to adversarial input.